Variable Interaction Networks

[mkommend]

This ticket tracks the necessary work for the automated generation and visualization of variable interaction networks in HeuristicLab. A brief description of desired functionality can be found in the attached word document.

[mkommend]

r11698: Added empty plugin for variable interaction networks.

[arapeanu]

r12198: Implemented Create Target Variation Experiment and added view skeleton for the variable interaction network

[arapeanu]

r12229: VariableInteractionNetworkView - calculates adjacency matrix from mean impacts between input and target variables

[arapeanu]

r12230: Updated project files (removed .resx file)

[arapeanu]

r12263: Added adjacency matrix update by threshold and target variable functionality + node importance calculation (based on adjacency matrix and weighted by the R² training coefficient)

[arapeanu]

r12320: Added sliders & functionality, dialog box for setting the number of repetitions per batch + modified node importance weighting

[arapeanu]

r12321: Modified CreateVariableCombinations in order to exclude unchecked inputs from the target variables and generate the batches accordingly

[arapeanu]

r12460: Added SymbolicDataAnalysisVariableImpactsAnalyzer for computing the average relative variable impacts over a GP with SR run

[ascheibe]

r12462 fixed plugin dependencies

[arapeanu]

r12568: Fixed bug in SymbolicDataAnalysisVariableImpactsAnalyzer (the simplification before optimization returning an error) and in VariableInteractionNetworkView for computing the (correctly sorted) adjacency matrix and sorting the variable impacts matrix

[bburlacu]

r13664: Change the CreateTargetVariationExperimentDialog so that it creates combinations of all variables in the dataset.

[bburlacu]

r13665: Improve calculation of variable impacts in the analyzer

[bburlacu]

r13727: Added directed graph chart and layout class for the visualization of knowledge network graphs. Added RunCollectionVariableInteractionNetworkView and removed the old view.

[bburlacu]

r13728: Added VariableInteractionNetwork graph class. Small improvements to the impacts analyzer. Add license header and improve formatting in CreateTargetVariationExperiment.cs

[bburlacu]

r13772: Add cloning constructors for the VariableNetworkNode and JunctionNetworkNode classes.

[bburlacu]

r13773: Refactor the RunCollectionVariableInteractionNetworkView and add online impact calculation (optimized method inside the view performs faster than the equivalent method provided by the RegressionSolutionVariableImpactsCalculator).

[bburlacu]

r13788: Improve the CreateTargetVariationExperimentDialog to produce combinations of inputs and target according to a user-provided binomial coefficient.

[bburlacu]

r13789:

• Refactor RunCollectionVariableInteractionNetworkView improving functionality, modularity and code organisation.
• Small tweaks to the DirectedGraphChart and DirectedGraphChartMode

[bburlacu]

r13806: Remove TinySet.cs in favor of a more general method for generating k-combinations. Improve target variation experiment generation. Refactored code and avoided some corner case exceptions.

[bburlacu]

r13814: Improve the CreateTargetVariationExperimentDialog r13815: Remove minimize and maximize buttons from the CreateTargetVariationExperimentDialog form

[bburlacu]

r13821: Used localization when formatting strings. Add canceling in CreateTargetVariationExperimentDialog. Remove use of C# 4.6 language features.

[bburlacu]

r13835: Performance improvements in the SymbolicDataAnalysisVariableImpactsAnalyzer.

[bburlacu]

r13868: Fix bug in impacts analyzer (wrong estimation limits when trying to optimize constants)

[bburlacu]

r13874: Small refactoring.

[bburlacu]

r13893: Simplify and optimize code for cluster identification in ConstrainedForceDirectedLayout.cs. Introduce a TrackBar for adjusting network threshold in the RunCollectionVariableInteractionNetworkView. Minor improvements to the DirectedGraphChart (work in progress).

[gkronber]

I cannot compile the branch because of missing dependency "Cola". Have I missed something? Could you please fix this?

[bburlacu]

r13967: Provide binaries of the external libraries upon which this branch depends. r13968: Update project file.

[gkronber]

r14031: moved native dlls out of the folder to make sure that HL discovers the files

[bburlacu]

r14275: Clean up code and add comments in the ConstrainedForceDirectedLayout class. Minor changes to view and directed graph chart. Introduced an INetworkNode interface for more flexibility. Updated cola and adaptagrams dlls with latest changes from upstream.

[bburlacu]

r14283: Small refactor

[gkronber]

r14622:

• added a static method to create a network (as a DAG) from NMSE vector and variable impacts matrix
• added cycle detection
• added conversion of a network to graphviz
• added conversion of a network to its adjacency matrix

[gkronber]

r14630: introduced base class for variable network instance description and implemented GPR and linear variable networks as specific classes

[gkronber]

r14655: added method to calculate the difference of two variable networks

[bburlacu]

r15421: Sync with trunk + Minor refactor.

[bburlacu]

r15929: Rename branch and update references.

[bburlacu]

r16295: Refactor code (use HL impacts calculator instead of manually calculating impacts, various fixes and improvements)

[jzenisek]

r16497:

• added possibility to create simple networks (only one input var set per target var, i.e. without junction nodes)
• fixed minor enumeration bug
• enabled network view updates

[jzenisek]

r16498: adapted online calculation of variable impacts within VIN-view according to new interface of RegressionSolutionVariableImpactsCalculator.CalculateImpacts(..)

[gkronber]

r16864: updated to .NET 4.6.1 and new persistence backend for compatibility with current trunk

[gkronber]

r16962: Add a class which allows to calculate the variable relevance (using a permutation test) but allows to consider correlated variables.

[jzenisek]

r16966:

• splitted creation of networks with/without possible cycles
• included additional check to prevent errors within the impact calculation